ANUJ
DATA×AI×ENGINEERING
INITIALIZING SYSTEM
ANUJ MUNDU
RETURN TO SELECTED WORK
SOURCE CODELIVE DEMO
PROJECT 02•AI ENGINEERING•AI / ML

OmniVision DocIntel API™ — Document Forensics & ELA

Asynchronous FastAPI microservice for digital document forensics, OpenCV quality inspection & Error Level Analysis

Python 3.11+FastAPI AsyncOpenCVError Level AnalysisPydantic v2DockerPrometheusStreamlitSSE
P95 Latency
< 65ms
Asynchronous OpenCV & ELA forensic pipeline
Splice Detection
ELA Std > 18
Quantization compression residual analysis
Rate Limiting
60 req/min
In-memory token bucket defense
Container Size
< 200 MB
Multi-stage Alpine/Debian slim Docker image
// INTERACTIVE SYSTEM TELEMETRY & DIAGNOSTIC LAB
LIVE RTSP INFERENCE [640×640 INT8]
40.3 FPS
P95 LATENCY: 24.8ms
NMS IoU: 0.45
RUNTIME ENGINE & LATENCY BENCHMARK COMPARATOR

Empirical benchmark comparing INT8 Post-Training Quantized ONNX against vanilla TorchScript C++ tracing.

INFERENCE BATCH SIZE:
P95 Latency
24.8ms
Deterministic SLA
Throughput
40.3 FPS
Video streaming limit
RAM Footprint
14.2 MB
Model weight & graph
CPU Usage
38%
8-Core Edge node
Target: Sub-30ms budget on edge hardware✓ 3.1x Faster Than TorchScript

01 // SYSTEM OVERVIEW

OmniVision-DocIntel API™ is an enterprise-grade digital document forensics and quality assurance microservice. Engineered with asynchronous FastAPI and OpenCV, it provides automated defense against manipulated invoices, forged tax returns, and spliced receipts through Error Level Analysis (ELA), Laplacian blur variance screening, orientation auto-correction, and Prometheus telemetry.

02 // THE PROBLEM & ENGINEERING SIGNIFICANCE

The Core Challenge

FinTech, InsurTech, and compliance platforms receive thousands of unstandardized document scans daily. Bad actors exploit manual review fatigue by digitally altering financial totals, dates, and account numbers using photo-editing software.

Why This Matters

Standard OCR engines blindly read spliced text without assessing pixel integrity, leading to millions in fraudulent loan disbursements and insurance payouts.

Key Constraints:
  • •Detecting pixel-level copy-paste splices on compressed JPEG/PNG documents without requiring heavy GPU deep learning models.
  • •Filtering out blurred or severely tilted mobile camera scans before downstream OCR pipeline execution.
  • •Maintaining sub-100ms API response times under high burst concurrency with strict rate-limiting guardrails.

03 // DATA PIPELINE & PREPROCESSING

Input Format: Multipart image/PDF file uploads (PNG, JPEG, WebP, single-page PDF scans)Sample Volume: Synthetic and empirical enterprise invoice/receipt forensics benchmarks
Transformation Steps:
  • Streaming in-memory byte buffer decoding via Pillow and OpenCV (zero disk writes)
  • Color channel normalization and grayscale projection for gradient computation
  • Laplacian operator convolution to evaluate focus quality and camera blur
Cleaning Strategy: Automated orientation rectification via Hough transform line analysis and thresholded border padding.

04 // SYSTEM ARCHITECTURE & DATA FLOW

Client (REST / Streamlit) -> FastAPI Async Gateway -> Token-Bucket Rate Limiter -> OpenCV Blur & Skew Engine -> Error Level Analysis (ELA) Forensics -> Prometheus Telemetry -> JSON Audit Response.

STEP 01FastAPI · Uvicorn
FastAPI Async Gateway

Non-blocking ASGI server with Pydantic v2 schema validation, CORS security, and health probes.

STEP 02OpenCV (cv2) · NumPy
Computer Vision QA Engine

Measures Laplacian variance (blur detection) and computes histogram exposure levels.

STEP 03Digital Forensics · PIL
ELA Forensic Analyzer

Resaves images at known 95% JPEG quality and computes pixel-level absolute difference residuals.

STEP 04Prometheus Client
Prometheus Metrics Exporter

Exposes real-time request counts, latency histograms, and forensic flag rates.

05 // MODEL ENGINEERING & HYPERPARAMETERS

Base Architecture: Digital Forensic Residual Analysis (ELA) + Laplacian Variance Focus Estimator

Calibrated against multi-tier tampering benchmarks with varying JPEG compression ratios.

Hyperparameters & Training Dynamics:
  • • Blur Threshold: Var < 100.0
  • • ELA Scale Factor: 10x
  • • ELA Tamper Threshold: Std > 18.0
  • • Rate Limit: 60 RPM
Loss Function: N/A (Deterministic Computer Vision & Information Forensics)
Trade-off Rationale: Employed deterministic ELA and OpenCV transforms rather than multi-gigabyte CNNs, achieving 65ms CPU execution with zero GPU dependency.

06 // FAILURE ANALYSIS & ZERO-TRUST SAFEGUARDS

OBSERVED FAILURE MODES UNDER STRESS
  • • Heavily re-compressed documents (compressed > 5 times) where ELA differences flatten out.
  • • Extremely low-resolution scans (< 300x300 pixels) where Laplacian variance is naturally low.
Mitigation & Fallback: Combined multi-scale ELA with edge-frequency histogram analysis; returns informative WARNING flags rather than hard 400 errors.

07 // PRODUCTION DEPLOYMENT SPECS

Serving Framework
FastAPI Async ASGI + Streamlit Clinical Demonstration Suite
Containerization
Multi-stage Dockerfile built on python:3.11-slim, total image footprint under 190 MB
P95 SLA
< 65ms per single-page invoice evaluation
Throughput
180 requests/sec on dual-core CPU container

08 // ARCHITECTURAL DECISIONS & TRADE-OFFS

Built native Error Level Analysis (ELA) without heavy neural network dependencies.
Why: ELA runs in 25ms on lightweight CPUs without expensive GPU cloud instances, making it feasible for edge and low-cost SaaS deployments.
Alternative Discarded: Heavy deep learning segmentation models (U-Net).
Enforced strict in-memory stream processing.
Why: Prevents temporary image files from ever touching disk, guaranteeing compliance with enterprise data security and privacy mandates.
Alternative Discarded: Writing uploads to temporary /tmp folders.

09 // PLANNED IMPROVEMENTS & NEXT REVISIONS

  • →Add automated font-consistency analysis using localized OCR character geometry.
  • →Integrate EXIF metadata tampering and camera serial hash verification.
BACK TO ALL PROJECTS
INSPECT ON GITHUB